[hmellor]

Overview

Split from #60823: replaces the vendored MiniMax-M3 config and ports the MiniMax-M3 code to the upstream Transformers MiniMaxM3VLConfig schema.

Related PRs:

• [Chore] Port more vendored configs and processors to upstream Transformers #60823: the larger unvendoring PR this is split from; MiniMax-M3 has been dropped from it.
• [Chore] Use upstream Transformers DeepseekV4Config #61392: the same split for DeepSeek-V4.

Claims

• Registers upstream MiniMaxM3VLConfig for minimax_m3_vl in _CONFIG_REGISTRY, because the hub remote code uses the old schema.
• vllm/transformers_utils/configs/minimax_m3.py shrinks to MiniMaxM3MTPConfig, a subclass of upstream MiniMaxM3VLTextConfig for standalone MTP checkpoints (model_type: "minimax_m3_mtp"). These copy the main model's 60-entry per-layer lists, which upstream rejects for a 1-layer config, so the subclass keeps the leading entries the MTP layer reads. No public checkpoint uses this format, but private ones may.
• The NVIDIA and AMD model code, indexers and vision tower read the upstream fields:
  • layer_types / mlp_layer_types replace sparse_attention_config["sparse_attention_freq"] / moe_layer_freq.
  • index_* fields replace the rest of sparse_attention_config.
  • rope_parameters replaces rope_theta / partial_rotary_factor / rope_scaling.
  • The vision tower reads spatial_merge_size / temporal_patch_size directly.
• Upstream reports hidden_act="silu" and drops swiglu_beta, so the swigluoai check is removed and beta is fixed at 1.0, the value in every checkpoint and in the upstream reference model. sparse_init_block and sparse_score_type fall back to the indexer's defaults (0, "max"), which also match every checkpoint.
• Fix: re-saved checkpoints in the upstream schema (e.g. lukealonso/MiniMax-M3-NVFP4) failed with "Unsupported activation: silu"; they now build.
• No numerics change: greedy outputs are bit-identical to main, and GSM8k and GPQA Diamond match.
• Open MiniMax-M3 PRs that add reads of the old field names will need a rebase once this lands: [MiniMax-M3 Performance] Prefer CuTe for MiniMax M3 FP8 indexer prefill #60930, [CI][MiniMax-M3] Serving features compatibility matrix #61370, [ROCm][PERF] Enable MiniMax-M3 Indexer CP #57909, [ROCm][Perf][MiniMax-M3] Fused sparse-layer decode via AITER #59653, [ROCm][Model][MiniMax-M3] Context-parallel sparse lightning indexer (Triton, opt-in) #57832, [ROCm][MiniMax-M3] Integrate ATOM mono with shared weights and batch eligibility #60528.

Validation

nvidia/MiniMax-M3-NVFP4 on 8x B200, main (b965ee564d) on GPUs 0-3 vs this branch on GPUs 4-7, TP4 each, transformers 5.19.0, FlashInfer autotuning off.

Check	main	This PR
GSM8k, 5-shot, 1319 questions, 2 runs	0.875, 0.870	0.870, 0.871
GPQA Diamond, 8 repeats	0.686	0.691
Greedy decode, 5 prompts (up to ~56k chars) x 64 tokens		bit-identical logprobs

Config and meta-device model dumps with the transformers-upgrade skill's scripts, for MiniMaxAI/MiniMax-M3, MiniMaxAI/MiniMax-M3-MXFP8, nvidia/MiniMax-M3-NVFP4 and amd/MiniMax-M3-MXFP4: 0 diffs in parameters and module attributes, for both the main model and the MTP draft (--speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'), including a draft loaded from a synthetic standalone MTP checkpoint in the old schema. lukealonso/MiniMax-M3-NVFP4 goes from the error above to building. The config diffs are only the schema renames. MTP was not served: no public MiniMax-M3 checkpoint ships MTP weights. The ROCm path was not served either; it is covered by the meta-device build and test_fused_shared_expert.py, which builds the AMD MiniMaxM3Model.

vllm serve nvidia/MiniMax-M3-NVFP4 --trust-remote-code --tensor-parallel-size 4 \
  --max-model-len 65536 --reasoning-parser minimax_m3 \
  --kernel-config '{"enable_flashinfer_autotune": false}'
python tests/evals/gsm8k/gsm8k_eval.py --port <port> --num-shots 5 --max-tokens 512
OPENAI_API_KEY=EMPTY python tests/evals/gpt_oss/run_gpqa_eval.py --model nvidia/MiniMax-M3-NVFP4 \
  --reasoning-effort high --base-url http://localhost:<port>/v1 --n-threads 128 --max-output-tokens 32768

pytest tests/kernels/attention/test_minimax_m3.py tests/model_executor/layers/test_fused_shared_expert.py
# main: 1 failed (test_fmha_sm100_indexer_matches_reference), 223 passed; this PR: 224 passed
pytest tests/transformers_utils/test_minimax_m3_config.py  # 1 passed
pytest tests/transformers_utils tests/config tests/models/test_registry.py tests/test_config.py
# main and this PR: the same 11 failed, 1237 passed (GPUs hidden for this run)

---

Pull Request Checklist

• I used vLLM's /pr-checklist skill. (Mandatory for agents, optional for humans).

• AI assistance was used during the creation of this PR.

• Design Fit: Minimizes impact on core components, reuses existing functionality, and justifies added complexity.

• Testing and Validation: Validates the change and ensures any added tests are meaningful and reliable, with CI coverage or documented CI resource constraints and validation performed outside CI.

• Code Quality and Style: Keeps code and comments clear and concise, and updates relevant documentation and examples.

• Pull Request Contents: Includes a brief summary and relevant links, supports claims with evidence, explains root causes and implementation trade-offs, and follows the contributing guide.

🤖 Generated with Claude Code

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @hmellor.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork